cui-llama.rn 1.5.0 → 1.6.0

package/android/src/main/java/com/rnllama/LlamaContext.java

@@ -1,645 +1,645 @@
-package com.rnllama;
-
-import com.facebook.react.bridge.Arguments;
-import com.facebook.react.bridge.WritableArray;
-import com.facebook.react.bridge.WritableMap;
-import com.facebook.react.bridge.ReadableMap;
-import com.facebook.react.bridge.ReadableArray;
-import com.facebook.react.bridge.ReactApplicationContext;
-import com.facebook.react.modules.core.DeviceEventManagerModule;
-
-import android.util.Log;
-import android.os.Build;
-import android.os.ParcelFileDescriptor;
-import android.net.Uri;
-import android.content.Intent;
-import android.content.res.AssetManager;
-
-import java.lang.StringBuilder;
-import java.io.BufferedReader;
-import java.io.FileReader;
-import java.io.File;
-import java.io.IOException;
-import java.io.InputStream;
-import java.io.FileInputStream;
-
-public class LlamaContext {
-  public static final String NAME = "RNLlamaContext";
-
-  private static String loadedLibrary = "";
-
-  private static class NativeLogCallback {
-    DeviceEventManagerModule.RCTDeviceEventEmitter eventEmitter;
-
-    public NativeLogCallback(ReactApplicationContext reactContext) {
-      this.eventEmitter = reactContext.getJSModule(DeviceEventManagerModule.RCTDeviceEventEmitter.class);
-    }
-
-    void emitNativeLog(String level, String text) {
-      WritableMap event = Arguments.createMap();
-      event.putString("level", level);
-      event.putString("text", text);
-      eventEmitter.emit("@RNLlama_onNativeLog", event);
-    }
-  }
-
-  static void toggleNativeLog(ReactApplicationContext reactContext, boolean enabled) {
-    if (LlamaContext.isArchNotSupported()) {
-      throw new IllegalStateException("Only 64-bit architectures are supported");
-    }
-    if (enabled) {
-      setupLog(new NativeLogCallback(reactContext));
-    } else {
-      unsetLog();
-    }
-  }
-
-  private int id;
-  private ReactApplicationContext reactContext;
-  private long context;
-  private WritableMap modelDetails;
-  private int jobId = -1;
-  private DeviceEventManagerModule.RCTDeviceEventEmitter eventEmitter;
-
-  private byte[] ggufHeader = {0x47, 0x47, 0x55, 0x46};
-
-  private boolean isGGUF(final String filepath, final ReactApplicationContext reactContext) {
-    byte[] fileHeader = new byte[4];
-    InputStream fis = null;
-    try {
-      if (filepath.startsWith("content")) {
-        Uri uri = Uri.parse(filepath);
-        reactContext.getApplicationContext().getContentResolver().takePersistableUriPermission(uri, Intent.FLAG_GRANT_READ_URI_PERMISSION);
-        fis = reactContext.getApplicationContext().getContentResolver().openInputStream(uri);
-      } else {
-        fis = new FileInputStream(filepath);
-      }
-
-      int bytesRead = fis.read(fileHeader);
-      if(bytesRead < 4) {
-        return false;
-      }
-      for(int i = 0; i < 4; i++){
-        if(fileHeader[i] != ggufHeader[i])
-          return false;
-      }
-      return true;
-    } catch (Exception e) {
-      Log.e(NAME, "Failed to check GGUF: " + e.getMessage());
-      return false;
-    }finally {
-      if (fis != null) {
-          try {
-              fis.close();
-          } catch (Exception e) {
-              Log.d(NAME, "Closing InputStream failed.");
-          }
-      }
-    }
-  }
-
-  public LlamaContext(int id, ReactApplicationContext reactContext, ReadableMap params) {
-    if (LlamaContext.isArchNotSupported()) {
-      throw new IllegalStateException("Only 64-bit architectures are supported");
-    }
-    if (!params.hasKey("model")) {
-      throw new IllegalArgumentException("Missing required parameter: model");
-    }
-
-    String modelName = params.getString("model");
-
-    if(!isGGUF(modelName, reactContext)) {
-      throw new IllegalArgumentException("File is not in GGUF format");
-    }
-
-    if ( modelName.startsWith("content://")) {
-    Uri uri = Uri.parse(modelName);
-      try {
-        ParcelFileDescriptor pfd = reactContext.getApplicationContext().getContentResolver().openFileDescriptor(uri, "r");
-        modelName = "" + pfd.getFd();
-      } catch (Exception e) {
-        Log.e(NAME, "Failed to convert to FD!");
-      }
-    }
-    
-    // Check if file has GGUF magic numbers
-    this.id = id;
-    eventEmitter = reactContext.getJSModule(DeviceEventManagerModule.RCTDeviceEventEmitter.class);
-    this.context = initContext(
-      // String model,
-      modelName,
-      // String chat_template,
-      params.hasKey("chat_template") ? params.getString("chat_template") : "",
-      // String reasoning_format,
-      params.hasKey("reasoning_format") ? params.getString("reasoning_format") : "none",
-      // boolean embedding,
-      params.hasKey("embedding") ? params.getBoolean("embedding") : false,
-      // int embd_normalize,
-      params.hasKey("embd_normalize") ? params.getInt("embd_normalize") : -1,
-      // int n_ctx,
-      params.hasKey("n_ctx") ? params.getInt("n_ctx") : 512,
-      // int n_batch,
-      params.hasKey("n_batch") ? params.getInt("n_batch") : 512,
-      // int n_ubatch,
-      params.hasKey("n_ubatch") ? params.getInt("n_ubatch") : 512,
-      // int n_threads,
-      params.hasKey("n_threads") ? params.getInt("n_threads") : 0,
-      // int n_gpu_layers, // TODO: Support this
-      params.hasKey("n_gpu_layers") ? params.getInt("n_gpu_layers") : 0,
-      // boolean flash_attn,
-      params.hasKey("flash_attn") ? params.getBoolean("flash_attn") : false,
-      // String cache_type_k,
-      params.hasKey("cache_type_k") ? params.getString("cache_type_k") : "f16",
-      // String cache_type_v,
-      params.hasKey("cache_type_v") ? params.getString("cache_type_v") : "f16",
-      // boolean use_mlock,
-      params.hasKey("use_mlock") ? params.getBoolean("use_mlock") : true,
-      // boolean use_mmap,
-      params.hasKey("use_mmap") ? params.getBoolean("use_mmap") : true,
-      //boolean vocab_only,
-      params.hasKey("vocab_only") ? params.getBoolean("vocab_only") : false,
-      // String lora,
-      params.hasKey("lora") ? params.getString("lora") : "",
-      // float lora_scaled,
-      params.hasKey("lora_scaled") ? (float) params.getDouble("lora_scaled") : 1.0f,
-      // ReadableArray lora_adapters,
-      params.hasKey("lora_list") ? params.getArray("lora_list") : null,
-      // float rope_freq_base,
-      params.hasKey("rope_freq_base") ? (float) params.getDouble("rope_freq_base") : 0.0f,
-      // float rope_freq_scale
-      params.hasKey("rope_freq_scale") ? (float) params.getDouble("rope_freq_scale") : 0.0f,
-      // int pooling_type,
-      params.hasKey("pooling_type") ? params.getInt("pooling_type") : -1,
-      // LoadProgressCallback load_progress_callback
-      params.hasKey("use_progress_callback") ? new LoadProgressCallback(this) : null
-    );
-    if (this.context == -1) {
-      throw new IllegalStateException("Failed to initialize context");
-    }
-    this.modelDetails = loadModelDetails(this.context);
-    this.reactContext = reactContext;
-  }
-
-  public void interruptLoad() {
-    interruptLoad(this.context);
-  }
-
-  public long getContext() {
-    return context;
-  }
-
-  public WritableMap getModelDetails() {
-    return modelDetails;
-  }
-
-  public String getLoadedLibrary() {
-    return loadedLibrary;
-  }
-
-  public WritableMap getFormattedChatWithJinja(String messages, String chatTemplate, ReadableMap params) {
-    String jsonSchema = params.hasKey("json_schema") ? params.getString("json_schema") : "";
-    String tools = params.hasKey("tools") ? params.getString("tools") : "";
-    Boolean parallelToolCalls = params.hasKey("parallel_tool_calls") ? params.getBoolean("parallel_tool_calls") : false;
-    String toolChoice = params.hasKey("tool_choice") ? params.getString("tool_choice") : "";
-    return getFormattedChatWithJinja(
-      this.context,
-      messages,
-      chatTemplate == null ? "" : chatTemplate,
-      jsonSchema,
-      tools,
-      parallelToolCalls,
-      toolChoice
-    );
-  }
-
-  public String getFormattedChat(String messages, String chatTemplate) {
-    return getFormattedChat(this.context, messages, chatTemplate == null ? "" : chatTemplate);
-  }
-
-  private void emitLoadProgress(int progress) {
-    WritableMap event = Arguments.createMap();
-    event.putInt("contextId", LlamaContext.this.id);
-    event.putInt("progress", progress);
-    eventEmitter.emit("@RNLlama_onInitContextProgress", event);
-  }
-
-  private static class LoadProgressCallback {
-    LlamaContext context;
-
-    public LoadProgressCallback(LlamaContext context) {
-      this.context = context;
-    }
-
-    void onLoadProgress(int progress) {
-      context.emitLoadProgress(progress);
-    }
-  }
-
-  private void emitPartialCompletion(WritableMap tokenResult) {
-    WritableMap event = Arguments.createMap();
-    event.putInt("contextId", LlamaContext.this.id);
-    event.putMap("tokenResult", tokenResult);
-    eventEmitter.emit("@RNLlama_onToken", event);
-  }
-
-  private static class PartialCompletionCallback {
-    LlamaContext context;
-    boolean emitNeeded;
-
-    public PartialCompletionCallback(LlamaContext context, boolean emitNeeded) {
-      this.context = context;
-      this.emitNeeded = emitNeeded;
-    }
-
-    void onPartialCompletion(WritableMap tokenResult) {
-      if (!emitNeeded) return;
-      context.emitPartialCompletion(tokenResult);
-    }
-  }
-
-  public WritableMap loadSession(String path) {
-    if (path == null || path.isEmpty()) {
-      throw new IllegalArgumentException("File path is empty");
-    }
-    File file = new File(path);
-    if (!file.exists()) {
-      throw new IllegalArgumentException("File does not exist: " + path);
-    }
-    WritableMap result = loadSession(this.context, path);
-    if (result.hasKey("error")) {
-      throw new IllegalStateException(result.getString("error"));
-    }
-    return result;
-  }
-
-  public int saveSession(String path, int size) {
-    if (path == null || path.isEmpty()) {
-      throw new IllegalArgumentException("File path is empty");
-    }
-    return saveSession(this.context, path, size);
-  }
-
-  public WritableMap completion(ReadableMap params) {
-    if (!params.hasKey("prompt")) {
-      throw new IllegalArgumentException("Missing required parameter: prompt");
-    }
-
-    double[][] logit_bias = new double[0][0];
-    if (params.hasKey("logit_bias")) {
-      ReadableArray logit_bias_array = params.getArray("logit_bias");
-      logit_bias = new double[logit_bias_array.size()][];
-      for (int i = 0; i < logit_bias_array.size(); i++) {
-        ReadableArray logit_bias_row = logit_bias_array.getArray(i);
-        logit_bias[i] = new double[logit_bias_row.size()];
-        for (int j = 0; j < logit_bias_row.size(); j++) {
-          logit_bias[i][j] = logit_bias_row.getDouble(j);
-        }
-      }
-    }
-
-    WritableMap result = doCompletion(
-      this.context,
-      // String prompt,
-      params.getString("prompt"),
-      // int chat_format,
-      params.hasKey("chat_format") ? params.getInt("chat_format") : 0,
-      // String grammar,
-      params.hasKey("grammar") ? params.getString("grammar") : "",
-      // String json_schema,
-      params.hasKey("json_schema") ? params.getString("json_schema") : "",
-      // boolean grammar_lazy,
-      params.hasKey("grammar_lazy") ? params.getBoolean("grammar_lazy") : false,
-      // ReadableArray grammar_triggers,
-      params.hasKey("grammar_triggers") ? params.getArray("grammar_triggers") : null,
-      // ReadableArray preserved_tokens,
-      params.hasKey("preserved_tokens") ? params.getArray("preserved_tokens") : null,
-      // float temperature,
-      params.hasKey("temperature") ? (float) params.getDouble("temperature") : 0.7f,
-      // int n_threads,
-      params.hasKey("n_threads") ? params.getInt("n_threads") : 0,
-      // int n_predict,
-      params.hasKey("n_predict") ? params.getInt("n_predict") : -1,
-      // int n_probs,
-      params.hasKey("n_probs") ? params.getInt("n_probs") : 0,
-      // int penalty_last_n,
-      params.hasKey("penalty_last_n") ? params.getInt("penalty_last_n") : 64,
-      // float penalty_repeat,
-      params.hasKey("penalty_repeat") ? (float) params.getDouble("penalty_repeat") : 1.00f,
-      // float penalty_freq,
-      params.hasKey("penalty_freq") ? (float) params.getDouble("penalty_freq") : 0.00f,
-      // float penalty_present,
-      params.hasKey("penalty_present") ? (float) params.getDouble("penalty_present") : 0.00f,
-      // float mirostat,
-      params.hasKey("mirostat") ? (float) params.getDouble("mirostat") : 0.00f,
-      // float mirostat_tau,
-      params.hasKey("mirostat_tau") ? (float) params.getDouble("mirostat_tau") : 5.00f,
-      // float mirostat_eta,
-      params.hasKey("mirostat_eta") ? (float) params.getDouble("mirostat_eta") : 0.10f,
-      // int top_k,
-      params.hasKey("top_k") ? params.getInt("top_k") : 40,
-      // float top_p,
-      params.hasKey("top_p") ? (float) params.getDouble("top_p") : 0.95f,
-      // float min_p,
-      params.hasKey("min_p") ? (float) params.getDouble("min_p") : 0.05f,
-      // float xtc_threshold,
-      params.hasKey("xtc_threshold") ? (float) params.getDouble("xtc_threshold") : 0.00f,
-      // float xtc_probability,
-      params.hasKey("xtc_probability") ? (float) params.getDouble("xtc_probability") : 0.00f,
-      // float typical_p,
-      params.hasKey("typical_p") ? (float) params.getDouble("typical_p") : 1.00f,
-      // int seed,
-      params.hasKey("seed") ? params.getInt("seed") : -1,
-      // String[] stop,
-      params.hasKey("stop") ? params.getArray("stop").toArrayList().toArray(new String[0]) : new String[0],
-      // boolean ignore_eos,
-      params.hasKey("ignore_eos") ? params.getBoolean("ignore_eos") : false,
-      // double[][] logit_bias,
-      logit_bias,
-      // float dry_multiplier,
-      params.hasKey("dry_multiplier") ? (float) params.getDouble("dry_multiplier") : 0.00f,
-      // float dry_base,
-      params.hasKey("dry_base") ? (float) params.getDouble("dry_base") : 1.75f,
-      // int dry_allowed_length,
-      params.hasKey("dry_allowed_length") ? params.getInt("dry_allowed_length") : 2,
-      // int dry_penalty_last_n,
-      params.hasKey("dry_penalty_last_n") ? params.getInt("dry_penalty_last_n") : -1,
-      // float top_n_sigma,
-      params.hasKey("top_n_sigma") ? (float) params.getDouble("top_n_sigma") : -1.0f,
-      // String[] dry_sequence_breakers, when undef, we use the default definition from common.h
-      params.hasKey("dry_sequence_breakers") ? params.getArray("dry_sequence_breakers").toArrayList().toArray(new String[0]) : new String[]{"\n", ":", "\"", "*"},
-      // PartialCompletionCallback partial_completion_callback
-      new PartialCompletionCallback(
-        this,
-        params.hasKey("emit_partial_completion") ? params.getBoolean("emit_partial_completion") : false
-      )
-    );
-    if (result.hasKey("error")) {
-      throw new IllegalStateException(result.getString("error"));
-    }
-    return result;
-  }
-
-  public void stopCompletion() {
-    stopCompletion(this.context);
-  }
-
-  public boolean isPredicting() {
-    return isPredicting(this.context);
-  }
-
-  public WritableMap tokenize(String text) {
-    WritableMap result = Arguments.createMap();
-    result.putArray("tokens", tokenize(this.context, text));
-    return result;
-  }
-
-  public String detokenize(ReadableArray tokens) {
-    int[] toks = new int[tokens.size()];
-    for (int i = 0; i < tokens.size(); i++) {
-      toks[i] = (int) tokens.getDouble(i);
-    }
-    return detokenize(this.context, toks);
-  }
-
-  public WritableMap getEmbedding(String text, ReadableMap params) {
-    if (isEmbeddingEnabled(this.context) == false) {
-      throw new IllegalStateException("Embedding is not enabled");
-    }
-    WritableMap result = embedding(
-      this.context,
-      text,
-      // int embd_normalize,
-      params.hasKey("embd_normalize") ? params.getInt("embd_normalize") : -1
-    );
-    if (result.hasKey("error")) {
-      throw new IllegalStateException(result.getString("error"));
-    }
-    return result;
-  }
-
-  public String bench(int pp, int tg, int pl, int nr) {
-    return bench(this.context, pp, tg, pl, nr);
-  }
-
-  public int applyLoraAdapters(ReadableArray loraAdapters) {
-    int result = applyLoraAdapters(this.context, loraAdapters);
-    if (result != 0) {
-      throw new IllegalStateException("Failed to apply lora adapters");
-    }
-    return result;
-  }
-
-  public void removeLoraAdapters() {
-    removeLoraAdapters(this.context);
-  }
-
-  public WritableArray getLoadedLoraAdapters() {
-    return getLoadedLoraAdapters(this.context);
-  }
-
-  public void release() {
-    freeContext(context);
-  }
-
-  static {
-    Log.d(NAME, "Primary ABI: " + Build.SUPPORTED_ABIS[0]);
-
-    String cpuFeatures = LlamaContext.getCpuFeatures();
-    Log.d(NAME, "CPU features: " + cpuFeatures);
-    boolean hasFp16 = cpuFeatures.contains("fp16") || cpuFeatures.contains("fphp");
-    boolean hasDotProd = cpuFeatures.contains("dotprod") || cpuFeatures.contains("asimddp");
-    boolean hasSve = cpuFeatures.contains("sve");
-    boolean hasI8mm = cpuFeatures.contains("i8mm");
-    boolean isAtLeastArmV82 = cpuFeatures.contains("asimd") && cpuFeatures.contains("crc32") && cpuFeatures.contains("aes");
-    boolean isAtLeastArmV84 = cpuFeatures.contains("dcpop") && cpuFeatures.contains("uscat");
-    Log.d(NAME, "- hasFp16: " + hasFp16);
-    Log.d(NAME, "- hasDotProd: " + hasDotProd);
-    Log.d(NAME, "- hasSve: " + hasSve);
-    Log.d(NAME, "- hasI8mm: " + hasI8mm);
-    Log.d(NAME, "- isAtLeastArmV82: " + isAtLeastArmV82);
-    Log.d(NAME, "- isAtLeastArmV84: " + isAtLeastArmV84);
-
-    // TODO: Add runtime check for cpu features
-    if (LlamaContext.isArm64V8a()) {
-      if (hasDotProd && hasI8mm) {
-        Log.d(NAME, "Loading librnllama_v8_2_dotprod_i8mm.so");
-        System.loadLibrary("rnllama_v8_2_dotprod_i8mm");
-        loadedLibrary = "rnllama_v8_2_dotprod_i8mm";
-      } else if (hasDotProd) {
-        Log.d(NAME, "Loading librnllama_v8_2_dotprod.so");
-        System.loadLibrary("rnllama_v8_2_dotprod");
-        loadedLibrary = "rnllama_v8_2_dotprod";
-      } else if (hasI8mm) {
-        Log.d(NAME, "Loading librnllama_v8_2_i8mm.so");
-        System.loadLibrary("rnllama_v8_2_i8mm");
-        loadedLibrary = "rnllama_v8_2_i8mm";
-      } else if (hasFp16) {
-        Log.d(NAME, "Loading librnllama_v8_2.so");
-        System.loadLibrary("rnllama_v8_2");
-        loadedLibrary = "rnllama_v8_2";
-      } else {
-        Log.d(NAME, "Loading default librnllama_v8.so");
-        System.loadLibrary("rnllama_v8");
-        loadedLibrary = "rnllama_v8";
-      }
-      //  Log.d(NAME, "Loading librnllama_v8_7.so with runtime feature detection");
-      //  System.loadLibrary("rnllama_v8_7");
-    } else if (LlamaContext.isX86_64()) {
-      Log.d(NAME, "Loading librnllama_x86_64.so");
-      System.loadLibrary("rnllama_x86_64");
-      loadedLibrary = "rnllama_x86_64";
-    } else {
-      Log.d(NAME, "ARM32 is not supported, skipping loading library");
-    }
-  }
-
-  public static boolean isArm64V8a() {
-    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a");
-  }
-
-  private static boolean isX86_64() {
-    return Build.SUPPORTED_ABIS[0].equals("x86_64");
-  }
-
-  private static boolean isArchNotSupported() {
-    return isArm64V8a() == false && isX86_64() == false;
-  }
-
-  public static String getCpuFeatures() {
-    File file = new File("/proc/cpuinfo");
-    StringBuilder stringBuilder = new StringBuilder();
-    try {
-      BufferedReader bufferedReader = new BufferedReader(new FileReader(file));
-      String line;
-      while ((line = bufferedReader.readLine()) != null) {
-        if (line.startsWith("Features")) {
-          stringBuilder.append(line);
-          break;
-        }
-      }
-      bufferedReader.close();
-      return stringBuilder.toString();
-    } catch (IOException e) {
-      Log.w(NAME, "Couldn't read /proc/cpuinfo", e);
-      return "";
-    }
-  }
-
-  public void emitModelProgressUpdate(int progress) {
-    WritableMap event = Arguments.createMap();
-    event.putInt("progress", progress);
-    eventEmitter.emit("@RNLlama_onInitContextProgress", event);
-  }
-
-  protected static native WritableMap modelInfo(
-    String model,
-    String[] skip
-  );
-  protected static native long initContext(
-    String model,
-    String chat_template,
-    String reasoning_format,
-    boolean embedding,
-    int embd_normalize,
-    int n_ctx,
-    int n_batch,
-    int n_ubatch,
-    int n_threads,
-    int n_gpu_layers, // TODO: Support this
-    boolean flash_attn,
-    String cache_type_k,
-    String cache_type_v,
-    boolean use_mlock,
-    boolean use_mmap,
-    boolean vocab_only,
-    String lora,
-    float lora_scaled,
-    ReadableArray lora_list,
-    float rope_freq_base,
-    float rope_freq_scale,
-    int pooling_type,
-    LoadProgressCallback load_progress_callback
-  );
-  protected static native void interruptLoad(long contextPtr);
-  protected static native WritableMap loadModelDetails(
-    long contextPtr
-  );
-  protected static native WritableMap getFormattedChatWithJinja(
-    long contextPtr,
-    String messages,
-    String chatTemplate,
-    String jsonSchema,
-    String tools,
-    boolean parallelToolCalls,
-    String toolChoice
-  );
-  protected static native String getFormattedChat(
-    long contextPtr,
-    String messages,
-    String chatTemplate
-  );
-  protected static native WritableMap loadSession(
-    long contextPtr,
-    String path
-  );
-  protected static native int saveSession(
-    long contextPtr,
-    String path,
-    int size
-  );
-  protected static native WritableMap doCompletion(
-    long context_ptr,
-    String prompt,
-    int chat_format,
-    String grammar,
-    String json_schema,
-    boolean grammar_lazy,
-    ReadableArray grammar_triggers,
-    ReadableArray preserved_tokens,
-    float temperature,
-    int n_threads,
-    int n_predict,
-    int n_probs,
-    int penalty_last_n,
-    float penalty_repeat,
-    float penalty_freq,
-    float penalty_present,
-    float mirostat,
-    float mirostat_tau,
-    float mirostat_eta,
-    int top_k,
-    float top_p,
-    float min_p,
-    float xtc_threshold,
-    float xtc_probability,
-    float typical_p,
-    int seed,
-    String[] stop,
-    boolean ignore_eos,
-    double[][] logit_bias,
-    float   dry_multiplier,
-    float   dry_base,
-    int dry_allowed_length,
-    int dry_penalty_last_n,
-    float top_n_sigma,
-    String[] dry_sequence_breakers,
-    PartialCompletionCallback partial_completion_callback
-  );
-  protected static native void stopCompletion(long contextPtr);
-  protected static native boolean isPredicting(long contextPtr);
-  protected static native WritableArray tokenize(long contextPtr, String text);
-  protected static native String detokenize(long contextPtr, int[] tokens);
-  protected static native boolean isEmbeddingEnabled(long contextPtr);
-  protected static native WritableMap embedding(
-    long contextPtr,
-    String text,
-    int embd_normalize
-  );
-  protected static native String bench(long contextPtr, int pp, int tg, int pl, int nr);
-  protected static native int applyLoraAdapters(long contextPtr, ReadableArray loraAdapters);
-  protected static native void removeLoraAdapters(long contextPtr);
-  protected static native WritableArray getLoadedLoraAdapters(long contextPtr);
-  protected static native void freeContext(long contextPtr);
-  protected static native void setupLog(NativeLogCallback logCallback);
-  protected static native void unsetLog();
-}
+package com.rnllama;
+
+import com.facebook.react.bridge.Arguments;
+import com.facebook.react.bridge.WritableArray;
+import com.facebook.react.bridge.WritableMap;
+import com.facebook.react.bridge.ReadableMap;
+import com.facebook.react.bridge.ReadableArray;
+import com.facebook.react.bridge.ReactApplicationContext;
+import com.facebook.react.modules.core.DeviceEventManagerModule;
+
+import android.util.Log;
+import android.os.Build;
+import android.os.ParcelFileDescriptor;
+import android.net.Uri;
+import android.content.Intent;
+import android.content.res.AssetManager;
+
+import java.lang.StringBuilder;
+import java.io.BufferedReader;
+import java.io.FileReader;
+import java.io.File;
+import java.io.IOException;
+import java.io.InputStream;
+import java.io.FileInputStream;
+
+public class LlamaContext {
+  public static final String NAME = "RNLlamaContext";
+
+  private static String loadedLibrary = "";
+
+  private static class NativeLogCallback {
+    DeviceEventManagerModule.RCTDeviceEventEmitter eventEmitter;
+
+    public NativeLogCallback(ReactApplicationContext reactContext) {
+      this.eventEmitter = reactContext.getJSModule(DeviceEventManagerModule.RCTDeviceEventEmitter.class);
+    }
+
+    void emitNativeLog(String level, String text) {
+      WritableMap event = Arguments.createMap();
+      event.putString("level", level);
+      event.putString("text", text);
+      eventEmitter.emit("@RNLlama_onNativeLog", event);
+    }
+  }
+
+  static void toggleNativeLog(ReactApplicationContext reactContext, boolean enabled) {
+    if (LlamaContext.isArchNotSupported()) {
+      throw new IllegalStateException("Only 64-bit architectures are supported");
+    }
+    if (enabled) {
+      setupLog(new NativeLogCallback(reactContext));
+    } else {
+      unsetLog();
+    }
+  }
+
+  private int id;
+  private ReactApplicationContext reactContext;
+  private long context;
+  private WritableMap modelDetails;
+  private int jobId = -1;
+  private DeviceEventManagerModule.RCTDeviceEventEmitter eventEmitter;
+
+  private byte[] ggufHeader = {0x47, 0x47, 0x55, 0x46};
+
+  private boolean isGGUF(final String filepath, final ReactApplicationContext reactContext) {
+    byte[] fileHeader = new byte[4];
+    InputStream fis = null;
+    try {
+      if (filepath.startsWith("content")) {
+        Uri uri = Uri.parse(filepath);
+        reactContext.getApplicationContext().getContentResolver().takePersistableUriPermission(uri, Intent.FLAG_GRANT_READ_URI_PERMISSION);
+        fis = reactContext.getApplicationContext().getContentResolver().openInputStream(uri);
+      } else {
+        fis = new FileInputStream(filepath);
+      }
+
+      int bytesRead = fis.read(fileHeader);
+      if(bytesRead < 4) {
+        return false;
+      }
+      for(int i = 0; i < 4; i++){
+        if(fileHeader[i] != ggufHeader[i])
+          return false;
+      }
+      return true;
+    } catch (Exception e) {
+      Log.e(NAME, "Failed to check GGUF: " + e.getMessage());
+      return false;
+    }finally {
+      if (fis != null) {
+          try {
+              fis.close();
+          } catch (Exception e) {
+              Log.d(NAME, "Closing InputStream failed.");
+          }
+      }
+    }
+  }
+
+  public LlamaContext(int id, ReactApplicationContext reactContext, ReadableMap params) {
+    if (LlamaContext.isArchNotSupported()) {
+      throw new IllegalStateException("Only 64-bit architectures are supported");
+    }
+    if (!params.hasKey("model")) {
+      throw new IllegalArgumentException("Missing required parameter: model");
+    }
+
+    String modelName = params.getString("model");
+
+    if(!isGGUF(modelName, reactContext)) {
+      throw new IllegalArgumentException("File is not in GGUF format");
+    }
+
+    if ( modelName.startsWith("content://")) {
+    Uri uri = Uri.parse(modelName);
+      try {
+        ParcelFileDescriptor pfd = reactContext.getApplicationContext().getContentResolver().openFileDescriptor(uri, "r");
+        modelName = "" + pfd.getFd();
+      } catch (Exception e) {
+        Log.e(NAME, "Failed to convert to FD!");
+      }
+    }
+    
+    // Check if file has GGUF magic numbers
+    this.id = id;
+    eventEmitter = reactContext.getJSModule(DeviceEventManagerModule.RCTDeviceEventEmitter.class);
+    this.context = initContext(
+      // String model,
+      modelName,
+      // String chat_template,
+      params.hasKey("chat_template") ? params.getString("chat_template") : "",
+      // String reasoning_format,
+      params.hasKey("reasoning_format") ? params.getString("reasoning_format") : "none",
+      // boolean embedding,
+      params.hasKey("embedding") ? params.getBoolean("embedding") : false,
+      // int embd_normalize,
+      params.hasKey("embd_normalize") ? params.getInt("embd_normalize") : -1,
+      // int n_ctx,
+      params.hasKey("n_ctx") ? params.getInt("n_ctx") : 512,
+      // int n_batch,
+      params.hasKey("n_batch") ? params.getInt("n_batch") : 512,
+      // int n_ubatch,
+      params.hasKey("n_ubatch") ? params.getInt("n_ubatch") : 512,
+      // int n_threads,
+      params.hasKey("n_threads") ? params.getInt("n_threads") : 0,
+      // int n_gpu_layers, // TODO: Support this
+      params.hasKey("n_gpu_layers") ? params.getInt("n_gpu_layers") : 0,
+      // boolean flash_attn,
+      params.hasKey("flash_attn") ? params.getBoolean("flash_attn") : false,
+      // String cache_type_k,
+      params.hasKey("cache_type_k") ? params.getString("cache_type_k") : "f16",
+      // String cache_type_v,
+      params.hasKey("cache_type_v") ? params.getString("cache_type_v") : "f16",
+      // boolean use_mlock,
+      params.hasKey("use_mlock") ? params.getBoolean("use_mlock") : true,
+      // boolean use_mmap,
+      params.hasKey("use_mmap") ? params.getBoolean("use_mmap") : true,
+      //boolean vocab_only,
+      params.hasKey("vocab_only") ? params.getBoolean("vocab_only") : false,
+      // String lora,
+      params.hasKey("lora") ? params.getString("lora") : "",
+      // float lora_scaled,
+      params.hasKey("lora_scaled") ? (float) params.getDouble("lora_scaled") : 1.0f,
+      // ReadableArray lora_adapters,
+      params.hasKey("lora_list") ? params.getArray("lora_list") : null,
+      // float rope_freq_base,
+      params.hasKey("rope_freq_base") ? (float) params.getDouble("rope_freq_base") : 0.0f,
+      // float rope_freq_scale
+      params.hasKey("rope_freq_scale") ? (float) params.getDouble("rope_freq_scale") : 0.0f,
+      // int pooling_type,
+      params.hasKey("pooling_type") ? params.getInt("pooling_type") : -1,
+      // LoadProgressCallback load_progress_callback
+      params.hasKey("use_progress_callback") ? new LoadProgressCallback(this) : null
+    );
+    if (this.context == -1) {
+      throw new IllegalStateException("Failed to initialize context");
+    }
+    this.modelDetails = loadModelDetails(this.context);
+    this.reactContext = reactContext;
+  }
+
+  public void interruptLoad() {
+    interruptLoad(this.context);
+  }
+
+  public long getContext() {
+    return context;
+  }
+
+  public WritableMap getModelDetails() {
+    return modelDetails;
+  }
+
+  public String getLoadedLibrary() {
+    return loadedLibrary;
+  }
+
+  public WritableMap getFormattedChatWithJinja(String messages, String chatTemplate, ReadableMap params) {
+    String jsonSchema = params.hasKey("json_schema") ? params.getString("json_schema") : "";
+    String tools = params.hasKey("tools") ? params.getString("tools") : "";
+    Boolean parallelToolCalls = params.hasKey("parallel_tool_calls") ? params.getBoolean("parallel_tool_calls") : false;
+    String toolChoice = params.hasKey("tool_choice") ? params.getString("tool_choice") : "";
+    return getFormattedChatWithJinja(
+      this.context,
+      messages,
+      chatTemplate == null ? "" : chatTemplate,
+      jsonSchema,
+      tools,
+      parallelToolCalls,
+      toolChoice
+    );
+  }
+
+  public String getFormattedChat(String messages, String chatTemplate) {
+    return getFormattedChat(this.context, messages, chatTemplate == null ? "" : chatTemplate);
+  }
+
+  private void emitLoadProgress(int progress) {
+    WritableMap event = Arguments.createMap();
+    event.putInt("contextId", LlamaContext.this.id);
+    event.putInt("progress", progress);
+    eventEmitter.emit("@RNLlama_onInitContextProgress", event);
+  }
+
+  private static class LoadProgressCallback {
+    LlamaContext context;
+
+    public LoadProgressCallback(LlamaContext context) {
+      this.context = context;
+    }
+
+    void onLoadProgress(int progress) {
+      context.emitLoadProgress(progress);
+    }
+  }
+
+  private void emitPartialCompletion(WritableMap tokenResult) {
+    WritableMap event = Arguments.createMap();
+    event.putInt("contextId", LlamaContext.this.id);
+    event.putMap("tokenResult", tokenResult);
+    eventEmitter.emit("@RNLlama_onToken", event);
+  }
+
+  private static class PartialCompletionCallback {
+    LlamaContext context;
+    boolean emitNeeded;
+
+    public PartialCompletionCallback(LlamaContext context, boolean emitNeeded) {
+      this.context = context;
+      this.emitNeeded = emitNeeded;
+    }
+
+    void onPartialCompletion(WritableMap tokenResult) {
+      if (!emitNeeded) return;
+      context.emitPartialCompletion(tokenResult);
+    }
+  }
+
+  public WritableMap loadSession(String path) {
+    if (path == null || path.isEmpty()) {
+      throw new IllegalArgumentException("File path is empty");
+    }
+    File file = new File(path);
+    if (!file.exists()) {
+      throw new IllegalArgumentException("File does not exist: " + path);
+    }
+    WritableMap result = loadSession(this.context, path);
+    if (result.hasKey("error")) {
+      throw new IllegalStateException(result.getString("error"));
+    }
+    return result;
+  }
+
+  public int saveSession(String path, int size) {
+    if (path == null || path.isEmpty()) {
+      throw new IllegalArgumentException("File path is empty");
+    }
+    return saveSession(this.context, path, size);
+  }
+
+  public WritableMap completion(ReadableMap params) {
+    if (!params.hasKey("prompt")) {
+      throw new IllegalArgumentException("Missing required parameter: prompt");
+    }
+
+    double[][] logit_bias = new double[0][0];
+    if (params.hasKey("logit_bias")) {
+      ReadableArray logit_bias_array = params.getArray("logit_bias");
+      logit_bias = new double[logit_bias_array.size()][];
+      for (int i = 0; i < logit_bias_array.size(); i++) {
+        ReadableArray logit_bias_row = logit_bias_array.getArray(i);
+        logit_bias[i] = new double[logit_bias_row.size()];
+        for (int j = 0; j < logit_bias_row.size(); j++) {
+          logit_bias[i][j] = logit_bias_row.getDouble(j);
+        }
+      }
+    }
+
+    WritableMap result = doCompletion(
+      this.context,
+      // String prompt,
+      params.getString("prompt"),
+      // int chat_format,
+      params.hasKey("chat_format") ? params.getInt("chat_format") : 0,
+      // String grammar,
+      params.hasKey("grammar") ? params.getString("grammar") : "",
+      // String json_schema,
+      params.hasKey("json_schema") ? params.getString("json_schema") : "",
+      // boolean grammar_lazy,
+      params.hasKey("grammar_lazy") ? params.getBoolean("grammar_lazy") : false,
+      // ReadableArray grammar_triggers,
+      params.hasKey("grammar_triggers") ? params.getArray("grammar_triggers") : null,
+      // ReadableArray preserved_tokens,
+      params.hasKey("preserved_tokens") ? params.getArray("preserved_tokens") : null,
+      // float temperature,
+      params.hasKey("temperature") ? (float) params.getDouble("temperature") : 0.7f,
+      // int n_threads,
+      params.hasKey("n_threads") ? params.getInt("n_threads") : 0,
+      // int n_predict,
+      params.hasKey("n_predict") ? params.getInt("n_predict") : -1,
+      // int n_probs,
+      params.hasKey("n_probs") ? params.getInt("n_probs") : 0,
+      // int penalty_last_n,
+      params.hasKey("penalty_last_n") ? params.getInt("penalty_last_n") : 64,
+      // float penalty_repeat,
+      params.hasKey("penalty_repeat") ? (float) params.getDouble("penalty_repeat") : 1.00f,
+      // float penalty_freq,
+      params.hasKey("penalty_freq") ? (float) params.getDouble("penalty_freq") : 0.00f,
+      // float penalty_present,
+      params.hasKey("penalty_present") ? (float) params.getDouble("penalty_present") : 0.00f,
+      // float mirostat,
+      params.hasKey("mirostat") ? (float) params.getDouble("mirostat") : 0.00f,
+      // float mirostat_tau,
+      params.hasKey("mirostat_tau") ? (float) params.getDouble("mirostat_tau") : 5.00f,
+      // float mirostat_eta,
+      params.hasKey("mirostat_eta") ? (float) params.getDouble("mirostat_eta") : 0.10f,
+      // int top_k,
+      params.hasKey("top_k") ? params.getInt("top_k") : 40,
+      // float top_p,
+      params.hasKey("top_p") ? (float) params.getDouble("top_p") : 0.95f,
+      // float min_p,
+      params.hasKey("min_p") ? (float) params.getDouble("min_p") : 0.05f,
+      // float xtc_threshold,
+      params.hasKey("xtc_threshold") ? (float) params.getDouble("xtc_threshold") : 0.00f,
+      // float xtc_probability,
+      params.hasKey("xtc_probability") ? (float) params.getDouble("xtc_probability") : 0.00f,
+      // float typical_p,
+      params.hasKey("typical_p") ? (float) params.getDouble("typical_p") : 1.00f,
+      // int seed,
+      params.hasKey("seed") ? params.getInt("seed") : -1,
+      // String[] stop,
+      params.hasKey("stop") ? params.getArray("stop").toArrayList().toArray(new String[0]) : new String[0],
+      // boolean ignore_eos,
+      params.hasKey("ignore_eos") ? params.getBoolean("ignore_eos") : false,
+      // double[][] logit_bias,
+      logit_bias,
+      // float dry_multiplier,
+      params.hasKey("dry_multiplier") ? (float) params.getDouble("dry_multiplier") : 0.00f,
+      // float dry_base,
+      params.hasKey("dry_base") ? (float) params.getDouble("dry_base") : 1.75f,
+      // int dry_allowed_length,
+      params.hasKey("dry_allowed_length") ? params.getInt("dry_allowed_length") : 2,
+      // int dry_penalty_last_n,
+      params.hasKey("dry_penalty_last_n") ? params.getInt("dry_penalty_last_n") : -1,
+      // float top_n_sigma,
+      params.hasKey("top_n_sigma") ? (float) params.getDouble("top_n_sigma") : -1.0f,
+      // String[] dry_sequence_breakers, when undef, we use the default definition from common.h
+      params.hasKey("dry_sequence_breakers") ? params.getArray("dry_sequence_breakers").toArrayList().toArray(new String[0]) : new String[]{"\n", ":", "\"", "*"},
+      // PartialCompletionCallback partial_completion_callback
+      new PartialCompletionCallback(
+        this,
+        params.hasKey("emit_partial_completion") ? params.getBoolean("emit_partial_completion") : false
+      )
+    );
+    if (result.hasKey("error")) {
+      throw new IllegalStateException(result.getString("error"));
+    }
+    return result;
+  }
+
+  public void stopCompletion() {
+    stopCompletion(this.context);
+  }
+
+  public boolean isPredicting() {
+    return isPredicting(this.context);
+  }
+
+  public WritableMap tokenize(String text) {
+    WritableMap result = Arguments.createMap();
+    result.putArray("tokens", tokenize(this.context, text));
+    return result;
+  }
+
+  public String detokenize(ReadableArray tokens) {
+    int[] toks = new int[tokens.size()];
+    for (int i = 0; i < tokens.size(); i++) {
+      toks[i] = (int) tokens.getDouble(i);
+    }
+    return detokenize(this.context, toks);
+  }
+
+  public WritableMap getEmbedding(String text, ReadableMap params) {
+    if (isEmbeddingEnabled(this.context) == false) {
+      throw new IllegalStateException("Embedding is not enabled");
+    }
+    WritableMap result = embedding(
+      this.context,
+      text,
+      // int embd_normalize,
+      params.hasKey("embd_normalize") ? params.getInt("embd_normalize") : -1
+    );
+    if (result.hasKey("error")) {
+      throw new IllegalStateException(result.getString("error"));
+    }
+    return result;
+  }
+
+  public String bench(int pp, int tg, int pl, int nr) {
+    return bench(this.context, pp, tg, pl, nr);
+  }
+
+  public int applyLoraAdapters(ReadableArray loraAdapters) {
+    int result = applyLoraAdapters(this.context, loraAdapters);
+    if (result != 0) {
+      throw new IllegalStateException("Failed to apply lora adapters");
+    }
+    return result;
+  }
+
+  public void removeLoraAdapters() {
+    removeLoraAdapters(this.context);
+  }
+
+  public WritableArray getLoadedLoraAdapters() {
+    return getLoadedLoraAdapters(this.context);
+  }
+
+  public void release() {
+    freeContext(context);
+  }
+
+  static {
+    Log.d(NAME, "Primary ABI: " + Build.SUPPORTED_ABIS[0]);
+
+    String cpuFeatures = LlamaContext.getCpuFeatures();
+    Log.d(NAME, "CPU features: " + cpuFeatures);
+    boolean hasFp16 = cpuFeatures.contains("fp16") || cpuFeatures.contains("fphp");
+    boolean hasDotProd = cpuFeatures.contains("dotprod") || cpuFeatures.contains("asimddp");
+    boolean hasSve = cpuFeatures.contains("sve");
+    boolean hasI8mm = cpuFeatures.contains("i8mm");
+    boolean isAtLeastArmV82 = cpuFeatures.contains("asimd") && cpuFeatures.contains("crc32") && cpuFeatures.contains("aes");
+    boolean isAtLeastArmV84 = cpuFeatures.contains("dcpop") && cpuFeatures.contains("uscat");
+    Log.d(NAME, "- hasFp16: " + hasFp16);
+    Log.d(NAME, "- hasDotProd: " + hasDotProd);
+    Log.d(NAME, "- hasSve: " + hasSve);
+    Log.d(NAME, "- hasI8mm: " + hasI8mm);
+    Log.d(NAME, "- isAtLeastArmV82: " + isAtLeastArmV82);
+    Log.d(NAME, "- isAtLeastArmV84: " + isAtLeastArmV84);
+
+    // TODO: Add runtime check for cpu features
+    if (LlamaContext.isArm64V8a()) {
+      if (hasDotProd && hasI8mm) {
+        Log.d(NAME, "Loading librnllama_v8_2_dotprod_i8mm.so");
+        System.loadLibrary("rnllama_v8_2_dotprod_i8mm");
+        loadedLibrary = "rnllama_v8_2_dotprod_i8mm";
+      } else if (hasDotProd) {
+        Log.d(NAME, "Loading librnllama_v8_2_dotprod.so");
+        System.loadLibrary("rnllama_v8_2_dotprod");
+        loadedLibrary = "rnllama_v8_2_dotprod";
+      } else if (hasI8mm) {
+        Log.d(NAME, "Loading librnllama_v8_2_i8mm.so");
+        System.loadLibrary("rnllama_v8_2_i8mm");
+        loadedLibrary = "rnllama_v8_2_i8mm";
+      } else if (hasFp16) {
+        Log.d(NAME, "Loading librnllama_v8_2.so");
+        System.loadLibrary("rnllama_v8_2");
+        loadedLibrary = "rnllama_v8_2";
+      } else {
+        Log.d(NAME, "Loading default librnllama_v8.so");
+        System.loadLibrary("rnllama_v8");
+        loadedLibrary = "rnllama_v8";
+      }
+      //  Log.d(NAME, "Loading librnllama_v8_7.so with runtime feature detection");
+      //  System.loadLibrary("rnllama_v8_7");
+    } else if (LlamaContext.isX86_64()) {
+      Log.d(NAME, "Loading librnllama_x86_64.so");
+      System.loadLibrary("rnllama_x86_64");
+      loadedLibrary = "rnllama_x86_64";
+    } else {
+      Log.d(NAME, "ARM32 is not supported, skipping loading library");
+    }
+  }
+
+  public static boolean isArm64V8a() {
+    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a");
+  }
+
+  private static boolean isX86_64() {
+    return Build.SUPPORTED_ABIS[0].equals("x86_64");
+  }
+
+  private static boolean isArchNotSupported() {
+    return isArm64V8a() == false && isX86_64() == false;
+  }
+
+  public static String getCpuFeatures() {
+    File file = new File("/proc/cpuinfo");
+    StringBuilder stringBuilder = new StringBuilder();
+    try {
+      BufferedReader bufferedReader = new BufferedReader(new FileReader(file));
+      String line;
+      while ((line = bufferedReader.readLine()) != null) {
+        if (line.startsWith("Features")) {
+          stringBuilder.append(line);
+          break;
+        }
+      }
+      bufferedReader.close();
+      return stringBuilder.toString();
+    } catch (IOException e) {
+      Log.w(NAME, "Couldn't read /proc/cpuinfo", e);
+      return "";
+    }
+  }
+
+  public void emitModelProgressUpdate(int progress) {
+    WritableMap event = Arguments.createMap();
+    event.putInt("progress", progress);
+    eventEmitter.emit("@RNLlama_onInitContextProgress", event);
+  }
+
+  protected static native WritableMap modelInfo(
+    String model,
+    String[] skip
+  );
+  protected static native long initContext(
+    String model,
+    String chat_template,
+    String reasoning_format,
+    boolean embedding,
+    int embd_normalize,
+    int n_ctx,
+    int n_batch,
+    int n_ubatch,
+    int n_threads,
+    int n_gpu_layers, // TODO: Support this
+    boolean flash_attn,
+    String cache_type_k,
+    String cache_type_v,
+    boolean use_mlock,
+    boolean use_mmap,
+    boolean vocab_only,
+    String lora,
+    float lora_scaled,
+    ReadableArray lora_list,
+    float rope_freq_base,
+    float rope_freq_scale,
+    int pooling_type,
+    LoadProgressCallback load_progress_callback
+  );
+  protected static native void interruptLoad(long contextPtr);
+  protected static native WritableMap loadModelDetails(
+    long contextPtr
+  );
+  protected static native WritableMap getFormattedChatWithJinja(
+    long contextPtr,
+    String messages,
+    String chatTemplate,
+    String jsonSchema,
+    String tools,
+    boolean parallelToolCalls,
+    String toolChoice
+  );
+  protected static native String getFormattedChat(
+    long contextPtr,
+    String messages,
+    String chatTemplate
+  );
+  protected static native WritableMap loadSession(
+    long contextPtr,
+    String path
+  );
+  protected static native int saveSession(
+    long contextPtr,
+    String path,
+    int size
+  );
+  protected static native WritableMap doCompletion(
+    long context_ptr,
+    String prompt,
+    int chat_format,
+    String grammar,
+    String json_schema,
+    boolean grammar_lazy,
+    ReadableArray grammar_triggers,
+    ReadableArray preserved_tokens,
+    float temperature,
+    int n_threads,
+    int n_predict,
+    int n_probs,
+    int penalty_last_n,
+    float penalty_repeat,
+    float penalty_freq,
+    float penalty_present,
+    float mirostat,
+    float mirostat_tau,
+    float mirostat_eta,
+    int top_k,
+    float top_p,
+    float min_p,
+    float xtc_threshold,
+    float xtc_probability,
+    float typical_p,
+    int seed,
+    String[] stop,
+    boolean ignore_eos,
+    double[][] logit_bias,
+    float   dry_multiplier,
+    float   dry_base,
+    int dry_allowed_length,
+    int dry_penalty_last_n,
+    float top_n_sigma,
+    String[] dry_sequence_breakers,
+    PartialCompletionCallback partial_completion_callback
+  );
+  protected static native void stopCompletion(long contextPtr);
+  protected static native boolean isPredicting(long contextPtr);
+  protected static native WritableArray tokenize(long contextPtr, String text);
+  protected static native String detokenize(long contextPtr, int[] tokens);
+  protected static native boolean isEmbeddingEnabled(long contextPtr);
+  protected static native WritableMap embedding(
+    long contextPtr,
+    String text,
+    int embd_normalize
+  );
+  protected static native String bench(long contextPtr, int pp, int tg, int pl, int nr);
+  protected static native int applyLoraAdapters(long contextPtr, ReadableArray loraAdapters);
+  protected static native void removeLoraAdapters(long contextPtr);
+  protected static native WritableArray getLoadedLoraAdapters(long contextPtr);
+  protected static native void freeContext(long contextPtr);
+  protected static native void setupLog(NativeLogCallback logCallback);
+  protected static native void unsetLog();
+}